Get-AQI in One shot-4 (GAOs-4)

To make the experimental results more accurate and the model more generalizable, we propose a new dataset GAOs-4. We used cell phone photography to manually collect 5700 environmental images at various times of the day in the Beijing area as our air quality dataset.
When collecting images, we try to avoid the interference of external factors on the images. For example, avoid capturing data in bad weather and low light, avoid direct sunlight to affect the representation of air quality in images, etc. The time, location, and AQI value of the acquisition are recorded at the same time as the image is acquired. The images were then divided into six categories based on the correspondence between AQI and air quality classes in Table II.
To improve the quality of the dataset and enhance the classification performance of the network, we carefully filter the collected dataset. We removed images that were blurred due to improper photography, were dim due to backlighting, and images where the sky and buildings were covered by nearby objects. We continuously filter and adjust the captured images to remove the redundant images with good air quality so that the images in each category are evenly distributed. We ended up with a high-quality environmental image dataset containing 3700 images. We divided them into 2960 training images and 740 validation images.

为提升实验结果的准确性与模型的泛化能力，我们提出了全新数据集GAOs-4。我们通过手机摄影，在北京地区分时段手动采集了5700张环境图像，以此作为空气质量数据集。
图像采集过程中，我们尽可能规避外界因素对成像的干扰：例如避免在恶劣天气与低光照环境下拍摄，避免直射阳光影响图像中空气质量特征的呈现等。采集图像的同时，我们同步记录了拍摄时间、拍摄地点与空气质量指数（AQI, Air Quality Index）数值。随后依据表二中的AQI与空气质量等级对应关系，将图像划分为六大类别。
为提升数据集质量、增强网络的分类性能，我们对采集得到的原始数据集进行了细致筛选：剔除了因拍摄不当导致模糊、因逆光导致过暗，以及天空与建筑被周边物体遮挡的图像。我们持续对采集图像进行筛选与调整，移除了空气质量良好类别的冗余样本，以确保各分类下的图像数量分布均衡。最终我们得到了包含3700张图像的高质量环境图像数据集，并将其划分为2960张训练图像与740张验证图像。